#include "lp_bld_type.h"
#include "lp_bld_logic.h"
#include "lp_bld_pack.h"
+#include "lp_bld_quad.h"
/*
/**
- * Initialize lp_sampler_static_state object with the gallium sampler
- * and texture state.
- * The former is considered to be static and the later dynamic.
+ * Initialize lp_sampler_static_texture_state object with the gallium
+ * texture/sampler_view state (this contains the parts which are
+ * considered static).
*/
void
-lp_sampler_static_state(struct lp_sampler_static_state *state,
- const struct pipe_sampler_view *view,
- const struct pipe_sampler_state *sampler)
+lp_sampler_static_texture_state(struct lp_static_texture_state *state,
+ const struct pipe_sampler_view *view)
{
- const struct pipe_resource *texture = view->texture;
+ const struct pipe_resource *texture;
memset(state, 0, sizeof *state);
- if(!texture)
+ if (!view || !view->texture)
return;
- if(!sampler)
+ texture = view->texture;
+
+ state->format = view->format;
+ state->swizzle_r = view->swizzle_r;
+ state->swizzle_g = view->swizzle_g;
+ state->swizzle_b = view->swizzle_b;
+ state->swizzle_a = view->swizzle_a;
+
+ state->target = texture->target;
+ state->pot_width = util_is_power_of_two(texture->width0);
+ state->pot_height = util_is_power_of_two(texture->height0);
+ state->pot_depth = util_is_power_of_two(texture->depth0);
+ state->level_zero_only = !view->u.tex.last_level;
+
+ /*
+ * the layer / element / level parameters are all either dynamic
+ * state or handled transparently wrt execution.
+ */
+}
+
+
+/**
+ * Initialize lp_sampler_static_sampler_state object with the gallium sampler
+ * state (this contains the parts which are considered static).
+ */
+void
+lp_sampler_static_sampler_state(struct lp_static_sampler_state *state,
+ const struct pipe_sampler_state *sampler)
+{
+ memset(state, 0, sizeof *state);
+
+ if (!sampler)
return;
/*
* regarding 1D/2D/3D/CUBE textures, wrap modes, etc.
*/
- state->format = view->format;
- state->swizzle_r = view->swizzle_r;
- state->swizzle_g = view->swizzle_g;
- state->swizzle_b = view->swizzle_b;
- state->swizzle_a = view->swizzle_a;
-
- state->target = texture->target;
- state->pot_width = util_is_power_of_two(texture->width0);
- state->pot_height = util_is_power_of_two(texture->height0);
- state->pot_depth = util_is_power_of_two(texture->depth0);
-
state->wrap_s = sampler->wrap_s;
state->wrap_t = sampler->wrap_t;
state->wrap_r = sampler->wrap_r;
state->min_img_filter = sampler->min_img_filter;
state->mag_img_filter = sampler->mag_img_filter;
- if (view->u.tex.last_level && sampler->max_lod > 0.0f) {
+ if (sampler->max_lod > 0.0f) {
state->min_mip_filter = sampler->min_mip_filter;
} else {
state->min_mip_filter = PIPE_TEX_MIPFILTER_NONE;
state->apply_min_lod = 1;
}
- if (sampler->max_lod < (float)view->u.tex.last_level) {
+ /*
+ * XXX this won't do anything with the mesa state tracker which always
+ * sets max_lod to not more than actually present mip maps...
+ */
+ if (sampler->max_lod < (PIPE_MAX_TEXTURE_LEVELS - 1)) {
state->apply_max_lod = 1;
}
}
}
state->normalized_coords = sampler->normalized_coords;
-
- /*
- * FIXME: Handle the remainder of pipe_sampler_view.
- */
}
*/
static LLVMValueRef
lp_build_rho(struct lp_build_sample_context *bld,
- unsigned unit,
+ unsigned texture_unit,
+ LLVMValueRef s,
+ LLVMValueRef t,
+ LLVMValueRef r,
+ LLVMValueRef cube_rho,
const struct lp_derivatives *derivs)
{
struct gallivm_state *gallivm = bld->gallivm;
- struct lp_build_context *int_size_bld = &bld->int_size_bld;
- struct lp_build_context *float_size_bld = &bld->float_size_bld;
+ struct lp_build_context *int_size_bld = &bld->int_size_in_bld;
+ struct lp_build_context *float_size_bld = &bld->float_size_in_bld;
struct lp_build_context *float_bld = &bld->float_bld;
struct lp_build_context *coord_bld = &bld->coord_bld;
struct lp_build_context *perquadf_bld = &bld->perquadf_bld;
- const LLVMValueRef *ddx_ddy = derivs->ddx_ddy;
const unsigned dims = bld->dims;
+ LLVMValueRef ddx_ddy[2];
LLVMBuilderRef builder = bld->gallivm->builder;
LLVMTypeRef i32t = LLVMInt32TypeInContext(bld->gallivm->context);
LLVMValueRef index0 = LLVMConstInt(i32t, 0, 0);
LLVMValueRef int_size, float_size;
LLVMValueRef rho;
LLVMValueRef first_level, first_level_vec;
- LLVMValueRef abs_ddx_ddy[2];
unsigned length = coord_bld->type.length;
unsigned num_quads = length / 4;
unsigned i;
LLVMValueRef i32undef = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
LLVMValueRef rho_xvec, rho_yvec;
- abs_ddx_ddy[0] = lp_build_abs(coord_bld, ddx_ddy[0]);
- if (dims > 2) {
- abs_ddx_ddy[1] = lp_build_abs(coord_bld, ddx_ddy[1]);
- }
- else {
- abs_ddx_ddy[1] = NULL;
- }
-
- if (dims == 1) {
- static const unsigned char swizzle1[] = {
- 0, LP_BLD_SWIZZLE_DONTCARE,
- LP_BLD_SWIZZLE_DONTCARE, LP_BLD_SWIZZLE_DONTCARE
- };
- static const unsigned char swizzle2[] = {
- 1, LP_BLD_SWIZZLE_DONTCARE,
- LP_BLD_SWIZZLE_DONTCARE, LP_BLD_SWIZZLE_DONTCARE
- };
- rho_xvec = lp_build_swizzle_aos(coord_bld, abs_ddx_ddy[0], swizzle1);
- rho_yvec = lp_build_swizzle_aos(coord_bld, abs_ddx_ddy[0], swizzle2);
- }
- else if (dims == 2) {
- static const unsigned char swizzle1[] = {
- 0, 2,
- LP_BLD_SWIZZLE_DONTCARE, LP_BLD_SWIZZLE_DONTCARE
- };
- static const unsigned char swizzle2[] = {
- 1, 3,
- LP_BLD_SWIZZLE_DONTCARE, LP_BLD_SWIZZLE_DONTCARE
- };
- rho_xvec = lp_build_swizzle_aos(coord_bld, abs_ddx_ddy[0], swizzle1);
- rho_yvec = lp_build_swizzle_aos(coord_bld, abs_ddx_ddy[0], swizzle2);
- }
- else {
- LLVMValueRef shuffles1[LP_MAX_VECTOR_LENGTH];
- LLVMValueRef shuffles2[LP_MAX_VECTOR_LENGTH];
- assert(dims == 3);
- for (i = 0; i < num_quads; i++) {
- shuffles1[4*i + 0] = lp_build_const_int32(gallivm, 4*i);
- shuffles1[4*i + 1] = lp_build_const_int32(gallivm, 4*i + 2);
- shuffles1[4*i + 2] = lp_build_const_int32(gallivm, length + 4*i);
- shuffles1[4*i + 3] = i32undef;
- shuffles2[4*i + 0] = lp_build_const_int32(gallivm, 4*i + 1);
- shuffles2[4*i + 1] = lp_build_const_int32(gallivm, 4*i + 3);
- shuffles2[4*i + 2] = lp_build_const_int32(gallivm, length + 4*i + 1);
- shuffles2[4*i + 3] = i32undef;
- }
- rho_xvec = LLVMBuildShuffleVector(builder, abs_ddx_ddy[0], abs_ddx_ddy[1],
- LLVMConstVector(shuffles1, length), "");
- rho_yvec = LLVMBuildShuffleVector(builder, abs_ddx_ddy[0], abs_ddx_ddy[1],
- LLVMConstVector(shuffles2, length), "");
- }
-
- rho_vec = lp_build_max(coord_bld, rho_xvec, rho_yvec);
+ /* Note that all simplified calculations will only work for isotropic filtering */
first_level = bld->dynamic_state->first_level(bld->dynamic_state,
- bld->gallivm, unit);
- first_level_vec = lp_build_broadcast_scalar(&bld->int_size_bld, first_level);
+ bld->gallivm, texture_unit);
+ first_level_vec = lp_build_broadcast_scalar(int_size_bld, first_level);
int_size = lp_build_minify(int_size_bld, bld->int_size, first_level_vec);
float_size = lp_build_int_to_float(float_size_bld, int_size);
- if (bld->coord_type.length > 4) {
- /* expand size to each quad */
+ if (cube_rho) {
+ LLVMValueRef cubesize;
+ LLVMValueRef index0 = lp_build_const_int32(gallivm, 0);
+ /*
+ * If we have derivs too then we have per-pixel cube_rho - doesn't matter
+ * though until we do per-pixel lod.
+ * Cube map code did already everything except size mul and per-quad extraction.
+ */
+ /* Could optimize this for single quad just skip the broadcast */
+ cubesize = lp_build_extract_broadcast(gallivm, bld->float_size_in_type,
+ coord_bld->type, float_size, index0);
+ rho_vec = lp_build_mul(coord_bld, cubesize, cube_rho);
+ rho = lp_build_pack_aos_scalars(bld->gallivm, coord_bld->type,
+ perquadf_bld->type, rho_vec, 0);
+ }
+ else if (derivs && !(bld->static_texture_state->target == PIPE_TEXTURE_CUBE)) {
+ LLVMValueRef ddmax[3];
+ for (i = 0; i < dims; i++) {
+ LLVMValueRef ddx, ddy;
+ LLVMValueRef floatdim;
+ LLVMValueRef indexi = lp_build_const_int32(gallivm, i);
+ ddx = lp_build_abs(coord_bld, derivs->ddx[i]);
+ ddy = lp_build_abs(coord_bld, derivs->ddy[i]);
+ ddmax[i] = lp_build_max(coord_bld, ddx, ddy);
+ floatdim = lp_build_extract_broadcast(gallivm, bld->float_size_in_type,
+ coord_bld->type, float_size, indexi);
+ ddmax[i] = lp_build_mul(coord_bld, floatdim, ddmax[i]);
+ }
+ rho_vec = ddmax[0];
if (dims > 1) {
- /* could use some broadcast_vector helper for this? */
- int num_quads = bld->coord_type.length / 4;
- LLVMValueRef src[LP_MAX_VECTOR_LENGTH/4];
- for (i = 0; i < num_quads; i++) {
- src[i] = float_size;
+ rho_vec = lp_build_max(coord_bld, rho_vec, ddmax[1]);
+ if (dims > 2) {
+ rho_vec = lp_build_max(coord_bld, rho_vec, ddmax[2]);
}
- float_size = lp_build_concat(bld->gallivm, src, float_size_bld->type, num_quads);
}
- else {
- float_size = lp_build_broadcast_scalar(coord_bld, float_size);
+ /*
+ * rho_vec now still contains per-pixel rho, convert to scalar per quad
+ * since we can't handle per-pixel rho/lod from now on (TODO).
+ */
+ rho = lp_build_pack_aos_scalars(bld->gallivm, coord_bld->type,
+ perquadf_bld->type, rho_vec, 0);
+ }
+ else {
+ /*
+ * This looks all a bit complex, but it's not that bad
+ * (the shuffle code makes it look worse than it is).
+ * Still, might not be ideal for all cases.
+ */
+ if (dims < 2) {
+ ddx_ddy[0] = lp_build_packed_ddx_ddy_onecoord(coord_bld, s);
+ }
+ else if (dims >= 2) {
+ ddx_ddy[0] = lp_build_packed_ddx_ddy_twocoord(coord_bld, s, t);
+ if (dims > 2) {
+ ddx_ddy[1] = lp_build_packed_ddx_ddy_onecoord(coord_bld, r);
+ }
}
- rho_vec = lp_build_mul(coord_bld, rho_vec, float_size);
- if (dims <= 1) {
- rho = rho_vec;
+ ddx_ddy[0] = lp_build_abs(coord_bld, ddx_ddy[0]);
+ if (dims > 2) {
+ ddx_ddy[1] = lp_build_abs(coord_bld, ddx_ddy[1]);
+ }
+
+ if (dims < 2) {
+ static const unsigned char swizzle1[] = { /* no-op swizzle */
+ 0, LP_BLD_SWIZZLE_DONTCARE,
+ LP_BLD_SWIZZLE_DONTCARE, LP_BLD_SWIZZLE_DONTCARE
+ };
+ static const unsigned char swizzle2[] = {
+ 2, LP_BLD_SWIZZLE_DONTCARE,
+ LP_BLD_SWIZZLE_DONTCARE, LP_BLD_SWIZZLE_DONTCARE
+ };
+ rho_xvec = lp_build_swizzle_aos(coord_bld, ddx_ddy[0], swizzle1);
+ rho_yvec = lp_build_swizzle_aos(coord_bld, ddx_ddy[0], swizzle2);
+ }
+ else if (dims == 2) {
+ static const unsigned char swizzle1[] = {
+ 0, 2,
+ LP_BLD_SWIZZLE_DONTCARE, LP_BLD_SWIZZLE_DONTCARE
+ };
+ static const unsigned char swizzle2[] = {
+ 1, 3,
+ LP_BLD_SWIZZLE_DONTCARE, LP_BLD_SWIZZLE_DONTCARE
+ };
+ rho_xvec = lp_build_swizzle_aos(coord_bld, ddx_ddy[0], swizzle1);
+ rho_yvec = lp_build_swizzle_aos(coord_bld, ddx_ddy[0], swizzle2);
}
else {
- if (dims >= 2) {
- static const unsigned char swizzle1[] = {
- 0, LP_BLD_SWIZZLE_DONTCARE,
- LP_BLD_SWIZZLE_DONTCARE, LP_BLD_SWIZZLE_DONTCARE
- };
- static const unsigned char swizzle2[] = {
- 1, LP_BLD_SWIZZLE_DONTCARE,
- LP_BLD_SWIZZLE_DONTCARE, LP_BLD_SWIZZLE_DONTCARE
- };
- LLVMValueRef rho_s, rho_t, rho_r;
-
- rho_s = lp_build_swizzle_aos(coord_bld, rho_vec, swizzle1);
- rho_t = lp_build_swizzle_aos(coord_bld, rho_vec, swizzle2);
-
- rho = lp_build_max(coord_bld, rho_s, rho_t);
-
- if (dims >= 3) {
- static const unsigned char swizzle3[] = {
- 2, LP_BLD_SWIZZLE_DONTCARE,
+ LLVMValueRef shuffles1[LP_MAX_VECTOR_LENGTH];
+ LLVMValueRef shuffles2[LP_MAX_VECTOR_LENGTH];
+ assert(dims == 3);
+ for (i = 0; i < num_quads; i++) {
+ shuffles1[4*i + 0] = lp_build_const_int32(gallivm, 4*i);
+ shuffles1[4*i + 1] = lp_build_const_int32(gallivm, 4*i + 2);
+ shuffles1[4*i + 2] = lp_build_const_int32(gallivm, length + 4*i);
+ shuffles1[4*i + 3] = i32undef;
+ shuffles2[4*i + 0] = lp_build_const_int32(gallivm, 4*i + 1);
+ shuffles2[4*i + 1] = lp_build_const_int32(gallivm, 4*i + 3);
+ shuffles2[4*i + 2] = lp_build_const_int32(gallivm, length + 4*i + 2);
+ shuffles2[4*i + 3] = i32undef;
+ }
+ rho_xvec = LLVMBuildShuffleVector(builder, ddx_ddy[0], ddx_ddy[1],
+ LLVMConstVector(shuffles1, length), "");
+ rho_yvec = LLVMBuildShuffleVector(builder, ddx_ddy[0], ddx_ddy[1],
+ LLVMConstVector(shuffles2, length), "");
+ }
+
+ rho_vec = lp_build_max(coord_bld, rho_xvec, rho_yvec);
+
+ if (bld->coord_type.length > 4) {
+ /* expand size to each quad */
+ if (dims > 1) {
+ /* could use some broadcast_vector helper for this? */
+ int num_quads = bld->coord_type.length / 4;
+ LLVMValueRef src[LP_MAX_VECTOR_LENGTH/4];
+ for (i = 0; i < num_quads; i++) {
+ src[i] = float_size;
+ }
+ float_size = lp_build_concat(bld->gallivm, src, float_size_bld->type, num_quads);
+ }
+ else {
+ float_size = lp_build_broadcast_scalar(coord_bld, float_size);
+ }
+ rho_vec = lp_build_mul(coord_bld, rho_vec, float_size);
+
+ if (dims <= 1) {
+ rho = rho_vec;
+ }
+ else {
+ if (dims >= 2) {
+ static const unsigned char swizzle1[] = {
+ 0, LP_BLD_SWIZZLE_DONTCARE,
+ LP_BLD_SWIZZLE_DONTCARE, LP_BLD_SWIZZLE_DONTCARE
+ };
+ static const unsigned char swizzle2[] = {
+ 1, LP_BLD_SWIZZLE_DONTCARE,
LP_BLD_SWIZZLE_DONTCARE, LP_BLD_SWIZZLE_DONTCARE
};
- rho_r = lp_build_swizzle_aos(coord_bld, rho_vec, swizzle3);
- rho = lp_build_max(coord_bld, rho, rho_r);
+ LLVMValueRef rho_s, rho_t, rho_r;
+
+ rho_s = lp_build_swizzle_aos(coord_bld, rho_vec, swizzle1);
+ rho_t = lp_build_swizzle_aos(coord_bld, rho_vec, swizzle2);
+
+ rho = lp_build_max(coord_bld, rho_s, rho_t);
+
+ if (dims >= 3) {
+ static const unsigned char swizzle3[] = {
+ 2, LP_BLD_SWIZZLE_DONTCARE,
+ LP_BLD_SWIZZLE_DONTCARE, LP_BLD_SWIZZLE_DONTCARE
+ };
+ rho_r = lp_build_swizzle_aos(coord_bld, rho_vec, swizzle3);
+ rho = lp_build_max(coord_bld, rho, rho_r);
+ }
}
}
- }
- rho = lp_build_pack_aos_scalars(bld->gallivm, coord_bld->type,
- perquadf_bld->type, rho);
- }
- else {
- if (dims <= 1) {
- rho_vec = LLVMBuildExtractElement(builder, rho_vec, index0, "");
- }
- rho_vec = lp_build_mul(float_size_bld, rho_vec, float_size);
-
- if (dims <= 1) {
- rho = rho_vec;
+ rho = lp_build_pack_aos_scalars(bld->gallivm, coord_bld->type,
+ perquadf_bld->type, rho, 0);
}
else {
- if (dims >= 2) {
- LLVMValueRef rho_s, rho_t, rho_r;
+ if (dims <= 1) {
+ rho_vec = LLVMBuildExtractElement(builder, rho_vec, index0, "");
+ }
+ rho_vec = lp_build_mul(float_size_bld, rho_vec, float_size);
- rho_s = LLVMBuildExtractElement(builder, rho_vec, index0, "");
- rho_t = LLVMBuildExtractElement(builder, rho_vec, index1, "");
+ if (dims <= 1) {
+ rho = rho_vec;
+ }
+ else {
+ if (dims >= 2) {
+ LLVMValueRef rho_s, rho_t, rho_r;
+
+ rho_s = LLVMBuildExtractElement(builder, rho_vec, index0, "");
+ rho_t = LLVMBuildExtractElement(builder, rho_vec, index1, "");
- rho = lp_build_max(float_bld, rho_s, rho_t);
+ rho = lp_build_max(float_bld, rho_s, rho_t);
- if (dims >= 3) {
- rho_r = LLVMBuildExtractElement(builder, rho_vec, index2, "");
- rho = lp_build_max(float_bld, rho, rho_r);
+ if (dims >= 3) {
+ rho_r = LLVMBuildExtractElement(builder, rho_vec, index2, "");
+ rho = lp_build_max(float_bld, rho, rho_r);
+ }
}
}
}
*/
void
lp_build_lod_selector(struct lp_build_sample_context *bld,
- unsigned unit,
+ unsigned texture_unit,
+ unsigned sampler_unit,
+ LLVMValueRef s,
+ LLVMValueRef t,
+ LLVMValueRef r,
+ LLVMValueRef cube_rho,
const struct lp_derivatives *derivs,
LLVMValueRef lod_bias, /* optional */
LLVMValueRef explicit_lod, /* optional */
*out_lod_ipart = bld->perquadi_bld.zero;
*out_lod_fpart = perquadf_bld->zero;
- if (bld->static_state->min_max_lod_equal) {
+ if (bld->static_sampler_state->min_max_lod_equal) {
/* User is forcing sampling from a particular mipmap level.
* This is hit during mipmap generation.
*/
LLVMValueRef min_lod =
- bld->dynamic_state->min_lod(bld->dynamic_state, bld->gallivm, unit);
+ bld->dynamic_state->min_lod(bld->dynamic_state,
+ bld->gallivm, sampler_unit);
lod = lp_build_broadcast_scalar(perquadf_bld, min_lod);
}
else {
if (explicit_lod) {
lod = lp_build_pack_aos_scalars(bld->gallivm, bld->coord_bld.type,
- perquadf_bld->type, explicit_lod);
+ perquadf_bld->type, explicit_lod, 0);
}
else {
LLVMValueRef rho;
- rho = lp_build_rho(bld, unit, derivs);
+ rho = lp_build_rho(bld, texture_unit, s, t, r, cube_rho, derivs);
/*
* Compute lod = log2(rho)
*/
if (!lod_bias &&
- !bld->static_state->lod_bias_non_zero &&
- !bld->static_state->apply_max_lod &&
- !bld->static_state->apply_min_lod) {
+ !bld->static_sampler_state->lod_bias_non_zero &&
+ !bld->static_sampler_state->apply_max_lod &&
+ !bld->static_sampler_state->apply_min_lod) {
/*
* Special case when there are no post-log2 adjustments, which
* saves instructions but keeping the integer and fractional lod
/* add shader lod bias */
if (lod_bias) {
lod_bias = lp_build_pack_aos_scalars(bld->gallivm, bld->coord_bld.type,
- perquadf_bld->type, lod_bias);
+ perquadf_bld->type, lod_bias, 0);
lod = LLVMBuildFAdd(builder, lod, lod_bias, "shader_lod_bias");
}
}
/* add sampler lod bias */
- if (bld->static_state->lod_bias_non_zero) {
+ if (bld->static_sampler_state->lod_bias_non_zero) {
LLVMValueRef sampler_lod_bias =
- bld->dynamic_state->lod_bias(bld->dynamic_state, bld->gallivm, unit);
+ bld->dynamic_state->lod_bias(bld->dynamic_state,
+ bld->gallivm, sampler_unit);
sampler_lod_bias = lp_build_broadcast_scalar(perquadf_bld,
sampler_lod_bias);
lod = LLVMBuildFAdd(builder, lod, sampler_lod_bias, "sampler_lod_bias");
}
/* clamp lod */
- if (bld->static_state->apply_max_lod) {
+ if (bld->static_sampler_state->apply_max_lod) {
LLVMValueRef max_lod =
- bld->dynamic_state->max_lod(bld->dynamic_state, bld->gallivm, unit);
+ bld->dynamic_state->max_lod(bld->dynamic_state,
+ bld->gallivm, sampler_unit);
max_lod = lp_build_broadcast_scalar(perquadf_bld, max_lod);
lod = lp_build_min(perquadf_bld, lod, max_lod);
}
- if (bld->static_state->apply_min_lod) {
+ if (bld->static_sampler_state->apply_min_lod) {
LLVMValueRef min_lod =
- bld->dynamic_state->min_lod(bld->dynamic_state, bld->gallivm, unit);
+ bld->dynamic_state->min_lod(bld->dynamic_state,
+ bld->gallivm, sampler_unit);
min_lod = lp_build_broadcast_scalar(perquadf_bld, min_lod);
lod = lp_build_max(perquadf_bld, lod, min_lod);
*/
void
lp_build_nearest_mip_level(struct lp_build_sample_context *bld,
- unsigned unit,
+ unsigned texture_unit,
LLVMValueRef lod_ipart,
LLVMValueRef *level_out)
{
LLVMValueRef first_level, last_level, level;
first_level = bld->dynamic_state->first_level(bld->dynamic_state,
- bld->gallivm, unit);
+ bld->gallivm, texture_unit);
last_level = bld->dynamic_state->last_level(bld->dynamic_state,
- bld->gallivm, unit);
+ bld->gallivm, texture_unit);
first_level = lp_build_broadcast_scalar(perquadi_bld, first_level);
last_level = lp_build_broadcast_scalar(perquadi_bld, last_level);
*/
void
lp_build_linear_mip_levels(struct lp_build_sample_context *bld,
- unsigned unit,
+ unsigned texture_unit,
LLVMValueRef lod_ipart,
LLVMValueRef *lod_fpart_inout,
LLVMValueRef *level0_out,
LLVMValueRef clamp_max;
first_level = bld->dynamic_state->first_level(bld->dynamic_state,
- bld->gallivm, unit);
+ bld->gallivm, texture_unit);
last_level = bld->dynamic_state->last_level(bld->dynamic_state,
- bld->gallivm, unit);
+ bld->gallivm, texture_unit);
first_level = lp_build_broadcast_scalar(perquadi_bld, first_level);
last_level = lp_build_broadcast_scalar(perquadi_bld, last_level);
*lod_fpart_inout = LLVMBuildSelect(builder, clamp_max,
perquadf_bld->zero, *lod_fpart_inout, "");
- lp_build_name(*level0_out, "sampler%u_miplevel0", unit);
- lp_build_name(*level1_out, "sampler%u_miplevel1", unit);
- lp_build_name(*lod_fpart_inout, "sampler%u_mipweight", unit);
+ lp_build_name(*level0_out, "texture%u_miplevel0", texture_unit);
+ lp_build_name(*level1_out, "texture%u_miplevel1", texture_unit);
+ lp_build_name(*lod_fpart_inout, "texture%u_mipweight", texture_unit);
}
/**
* Return pointer to a single mipmap level.
- * \param data_array array of pointers to mipmap levels
* \param level integer mipmap level
*/
LLVMValueRef
LLVMValueRef level)
{
LLVMBuilderRef builder = bld->gallivm->builder;
- LLVMValueRef indexes[2], data_ptr;
+ LLVMValueRef indexes[2], data_ptr, mip_offset;
indexes[0] = lp_build_const_int32(bld->gallivm, 0);
indexes[1] = level;
- data_ptr = LLVMBuildGEP(builder, bld->data_array, indexes, 2, "");
- data_ptr = LLVMBuildLoad(builder, data_ptr, "");
+ mip_offset = LLVMBuildGEP(builder, bld->mip_offsets, indexes, 2, "");
+ mip_offset = LLVMBuildLoad(builder, mip_offset, "");
+ data_ptr = LLVMBuildGEP(builder, bld->base_ptr, &mip_offset, 1, "");
return data_ptr;
}
+/**
+ * Return (per-pixel) offsets to mip levels.
+ * \param level integer mipmap level
+ */
+LLVMValueRef
+lp_build_get_mip_offsets(struct lp_build_sample_context *bld,
+ LLVMValueRef level)
+{
+ LLVMBuilderRef builder = bld->gallivm->builder;
+ LLVMValueRef indexes[2], offsets, offset1;
+
+ indexes[0] = lp_build_const_int32(bld->gallivm, 0);
+ if (bld->num_lods == 1) {
+ indexes[1] = level;
+ offset1 = LLVMBuildGEP(builder, bld->mip_offsets, indexes, 2, "");
+ offset1 = LLVMBuildLoad(builder, offset1, "");
+ offsets = lp_build_broadcast_scalar(&bld->int_coord_bld, offset1);
+ }
+ else if (bld->num_lods == bld->coord_bld.type.length / 4) {
+ unsigned i;
+
+ offsets = bld->int_coord_bld.undef;
+ for (i = 0; i < bld->num_lods; i++) {
+ LLVMValueRef indexi = lp_build_const_int32(bld->gallivm, i);
+ LLVMValueRef indexo = lp_build_const_int32(bld->gallivm, 4 * i);
+ indexes[1] = LLVMBuildExtractElement(builder, level, indexi, "");
+ offset1 = LLVMBuildGEP(builder, bld->mip_offsets, indexes, 2, "");
+ offset1 = LLVMBuildLoad(builder, offset1, "");
+ offsets = LLVMBuildInsertElement(builder, offsets, offset1, indexo, "");
+ }
+ offsets = lp_build_swizzle_scalar_aos(&bld->int_coord_bld, offsets, 0, 4);
+ }
+ else {
+ unsigned i;
+
+ assert (bld->num_lods == bld->coord_bld.type.length);
+
+ offsets = bld->int_coord_bld.undef;
+ for (i = 0; i < bld->num_lods; i++) {
+ LLVMValueRef indexi = lp_build_const_int32(bld->gallivm, i);
+ indexes[1] = LLVMBuildExtractElement(builder, level, indexi, "");
+ offset1 = LLVMBuildGEP(builder, bld->mip_offsets, indexes, 2, "");
+ offset1 = LLVMBuildLoad(builder, offset1, "");
+ offsets = LLVMBuildInsertElement(builder, offsets, offset1, indexi, "");
+ }
+ }
+ return offsets;
+}
+
/**
* Codegen equivalent for u_minify().
LLVMValueRef stride_array, LLVMValueRef level)
{
LLVMBuilderRef builder = bld->gallivm->builder;
- LLVMValueRef indexes[2], stride;
+ LLVMValueRef indexes[2], stride, stride1;
indexes[0] = lp_build_const_int32(bld->gallivm, 0);
- indexes[1] = level;
- stride = LLVMBuildGEP(builder, stride_array, indexes, 2, "");
- stride = LLVMBuildLoad(builder, stride, "");
- stride = lp_build_broadcast_scalar(&bld->int_coord_bld, stride);
+ if (bld->num_lods == 1) {
+ indexes[1] = level;
+ stride1 = LLVMBuildGEP(builder, stride_array, indexes, 2, "");
+ stride1 = LLVMBuildLoad(builder, stride1, "");
+ stride = lp_build_broadcast_scalar(&bld->int_coord_bld, stride1);
+ }
+ else if (bld->num_lods == bld->coord_bld.type.length / 4) {
+ LLVMValueRef stride1;
+ unsigned i;
+
+ stride = bld->int_coord_bld.undef;
+ for (i = 0; i < bld->num_lods; i++) {
+ LLVMValueRef indexi = lp_build_const_int32(bld->gallivm, i);
+ LLVMValueRef indexo = lp_build_const_int32(bld->gallivm, i);
+ indexes[1] = LLVMBuildExtractElement(builder, level, indexi, "");
+ stride1 = LLVMBuildGEP(builder, stride_array, indexes, 2, "");
+ stride1 = LLVMBuildLoad(builder, stride1, "");
+ stride = LLVMBuildInsertElement(builder, stride, stride1, indexo, "");
+ }
+ stride = lp_build_swizzle_scalar_aos(&bld->int_coord_bld, stride, 0, 4);
+ }
+ else {
+ LLVMValueRef stride1;
+ unsigned i;
+
+ assert (bld->num_lods == bld->coord_bld.type.length);
+
+ stride = bld->int_coord_bld.undef;
+ for (i = 0; i < bld->coord_bld.type.length; i++) {
+ LLVMValueRef indexi = lp_build_const_int32(bld->gallivm, i);
+ indexes[1] = LLVMBuildExtractElement(builder, level, indexi, "");
+ stride1 = LLVMBuildGEP(builder, stride_array, indexes, 2, "");
+ stride1 = LLVMBuildLoad(builder, stride1, "");
+ stride = LLVMBuildInsertElement(builder, stride, stride1, indexi, "");
+ }
+ }
return stride;
}
const unsigned dims = bld->dims;
LLVMValueRef ilevel_vec;
- ilevel_vec = lp_build_broadcast_scalar(&bld->int_size_bld, ilevel);
-
/*
* Compute width, height, depth at mipmap level 'ilevel'
*/
- *out_size = lp_build_minify(&bld->int_size_bld, bld->int_size, ilevel_vec);
+ if (bld->num_lods == 1) {
+ ilevel_vec = lp_build_broadcast_scalar(&bld->int_size_bld, ilevel);
+ *out_size = lp_build_minify(&bld->int_size_bld, bld->int_size, ilevel_vec);
+ }
+ else {
+ LLVMValueRef int_size_vec;
+ LLVMValueRef tmp[LP_MAX_VECTOR_LENGTH];
+ unsigned num_quads = bld->coord_bld.type.length / 4;
+ unsigned i;
+
+ if (bld->num_lods == num_quads) {
+ /*
+ * XXX: this should be #ifndef SANE_INSTRUCTION_SET.
+ * intel "forgot" the variable shift count instruction until avx2.
+ * A harmless 8x32 shift gets translated into 32 instructions
+ * (16 extracts, 8 scalar shifts, 8 inserts), llvm is apparently
+ * unable to recognize if there are really just 2 different shift
+ * count values. So do the shift 4-wide before expansion.
+ */
+ struct lp_build_context bld4;
+ struct lp_type type4;
+
+ type4 = bld->int_coord_bld.type;
+ type4.length = 4;
+
+ lp_build_context_init(&bld4, bld->gallivm, type4);
+
+ if (bld->dims == 1) {
+ assert(bld->int_size_in_bld.type.length == 1);
+ int_size_vec = lp_build_broadcast_scalar(&bld4,
+ bld->int_size);
+ }
+ else {
+ assert(bld->int_size_in_bld.type.length == 4);
+ int_size_vec = bld->int_size;
+ }
+
+ for (i = 0; i < num_quads; i++) {
+ LLVMValueRef ileveli;
+ LLVMValueRef indexi = lp_build_const_int32(bld->gallivm, i);
+
+ ileveli = lp_build_extract_broadcast(bld->gallivm,
+ bld->perquadi_bld.type,
+ bld4.type,
+ ilevel,
+ indexi);
+ tmp[i] = lp_build_minify(&bld4, int_size_vec, ileveli);
+ }
+ /*
+ * out_size is [w0, h0, d0, _, w1, h1, d1, _, ...] vector for dims > 1,
+ * [w0, w0, w0, w0, w1, w1, w1, w1, ...] otherwise.
+ */
+ *out_size = lp_build_concat(bld->gallivm,
+ tmp,
+ bld4.type,
+ num_quads);
+ }
+ else {
+ /* FIXME: this is terrible and results in _huge_ vector
+ * (for the dims > 1 case).
+ * Should refactor this (together with extract_image_sizes) and do
+ * something more useful. Could for instance if we have width,height
+ * with 4-wide vector pack all elements into a 8xi16 vector
+ * (on which we can still do useful math) instead of using a 16xi32
+ * vector.
+ * FIXME: some callers can't handle this yet.
+ * For dims == 1 this will create [w0, w1, w2, w3, ...] vector.
+ * For dims > 1 this will create [w0, h0, d0, _, w1, h1, d1, _, ...] vector.
+ */
+ assert(bld->num_lods == bld->coord_bld.type.length);
+ if (bld->dims == 1) {
+ assert(bld->int_size_bld.type.length == 1);
+ int_size_vec = lp_build_broadcast_scalar(&bld->int_coord_bld,
+ bld->int_size);
+ /* vector shift with variable shift count alert... */
+ *out_size = lp_build_minify(&bld->int_coord_bld, int_size_vec, ilevel);
+ }
+ else {
+ LLVMValueRef ilevel1;
+ for (i = 0; i < bld->num_lods; i++) {
+ LLVMValueRef indexi = lp_build_const_int32(bld->gallivm, i);
+ ilevel1 = lp_build_extract_broadcast(bld->gallivm, bld->int_coord_type,
+ bld->int_size_in_bld.type, ilevel, indexi);
+ tmp[i] = bld->int_size;
+ tmp[i] = lp_build_minify(&bld->int_size_in_bld, tmp[i], ilevel1);
+ }
+ int_size_vec = lp_build_concat(bld->gallivm,
+ tmp,
+ bld->int_size_in_bld.type,
+ bld->num_lods);
+ }
+ }
+ }
if (dims >= 2) {
*row_stride_vec = lp_build_get_level_stride_vec(bld,
bld->row_stride_array,
ilevel);
- if (dims == 3 || bld->static_state->target == PIPE_TEXTURE_CUBE) {
- *img_stride_vec = lp_build_get_level_stride_vec(bld,
- bld->img_stride_array,
- ilevel);
- }
+ }
+ if (dims == 3 ||
+ bld->static_texture_state->target == PIPE_TEXTURE_CUBE ||
+ bld->static_texture_state->target == PIPE_TEXTURE_1D_ARRAY ||
+ bld->static_texture_state->target == PIPE_TEXTURE_2D_ARRAY) {
+ *img_stride_vec = lp_build_get_level_stride_vec(bld,
+ bld->img_stride_array,
+ ilevel);
}
}
*/
void
lp_build_extract_image_sizes(struct lp_build_sample_context *bld,
- struct lp_type size_type,
+ struct lp_build_context *size_bld,
struct lp_type coord_type,
LLVMValueRef size,
LLVMValueRef *out_width,
{
const unsigned dims = bld->dims;
LLVMTypeRef i32t = LLVMInt32TypeInContext(bld->gallivm->context);
+ struct lp_type size_type = size_bld->type;
+
+ if (bld->num_lods == 1) {
+ *out_width = lp_build_extract_broadcast(bld->gallivm,
+ size_type,
+ coord_type,
+ size,
+ LLVMConstInt(i32t, 0, 0));
+ if (dims >= 2) {
+ *out_height = lp_build_extract_broadcast(bld->gallivm,
+ size_type,
+ coord_type,
+ size,
+ LLVMConstInt(i32t, 1, 0));
+ if (dims == 3) {
+ *out_depth = lp_build_extract_broadcast(bld->gallivm,
+ size_type,
+ coord_type,
+ size,
+ LLVMConstInt(i32t, 2, 0));
+ }
+ }
+ }
+ else {
+ unsigned num_quads = bld->coord_bld.type.length / 4;
- *out_width = lp_build_extract_broadcast(bld->gallivm,
- size_type,
- coord_type,
- size,
- LLVMConstInt(i32t, 0, 0));
- if (dims >= 2) {
- *out_height = lp_build_extract_broadcast(bld->gallivm,
- size_type,
- coord_type,
- size,
- LLVMConstInt(i32t, 1, 0));
- if (dims == 3) {
- *out_depth = lp_build_extract_broadcast(bld->gallivm,
- size_type,
- coord_type,
- size,
- LLVMConstInt(i32t, 2, 0));
+ if (dims == 1) {
+ *out_width = size;
+ }
+ else if (bld->num_lods == num_quads) {
+ *out_width = lp_build_swizzle_scalar_aos(size_bld, size, 0, 4);
+ if (dims >= 2) {
+ *out_height = lp_build_swizzle_scalar_aos(size_bld, size, 1, 4);
+ if (dims == 3) {
+ *out_depth = lp_build_swizzle_scalar_aos(size_bld, size, 2, 4);
+ }
+ }
+ }
+ else {
+ assert(bld->num_lods == bld->coord_type.length);
+ *out_width = lp_build_pack_aos_scalars(bld->gallivm, size_type,
+ coord_type, size, 0);
+ if (dims >= 2) {
+ *out_width = lp_build_pack_aos_scalars(bld->gallivm, size_type,
+ coord_type, size, 1);
+ if (dims == 3) {
+ *out_width = lp_build_pack_aos_scalars(bld->gallivm, size_type,
+ coord_type, size, 2);
+ }
+ }
}
}
}
LLVMValueRef depth;
lp_build_extract_image_sizes(bld,
- bld->float_size_type,
+ &bld->float_size_bld,
bld->coord_type,
flt_size,
&width,
LLVMValueRef r,
LLVMValueRef *face,
LLVMValueRef *face_s,
- LLVMValueRef *face_t)
+ LLVMValueRef *face_t,
+ LLVMValueRef *rho)
{
struct lp_build_context *coord_bld = &bld->coord_bld;
LLVMBuilderRef builder = bld->gallivm->builder;
struct gallivm_state *gallivm = bld->gallivm;
- LLVMValueRef rx, ry, rz;
- LLVMValueRef tmp[4], rxyz, arxyz;
+ LLVMValueRef si, ti, ri;
+ boolean implicit_derivs = TRUE;
+ boolean need_derivs = TRUE;
- /*
- * Use the average of the four pixel's texcoords to choose the face.
- * Slight simplification just calculate the sum, skip scaling.
- */
- tmp[0] = s;
- tmp[1] = t;
- tmp[2] = r;
- rxyz = lp_build_hadd_partial4(&bld->coord_bld, tmp, 3);
- arxyz = lp_build_abs(&bld->coord_bld, rxyz);
-
- if (coord_bld->type.length > 4) {
+ if (1 || coord_bld->type.length > 4) {
+ /*
+ * Do per-pixel face selection. We cannot however (as we used to do)
+ * simply calculate the derivs afterwards (which is very bogus for
+ * explicit derivs anyway) because the values would be "random" when
+ * not all pixels lie on the same face. Hence just transform the derivs
+ * (or rather only the dmax values), which works both for implicit and
+ * explicit derivatives and doesn't add much math (except need to
+ * calculate derivs for 3 instead of 2 coords and have a couple more selects
+ * but cuts some minor math elsewhere). The derivs don't need mirroring,
+ * just selection, since noone cares about the sign.
+ */
struct lp_build_context *cint_bld = &bld->int_coord_bld;
struct lp_type intctype = cint_bld->type;
- LLVMValueRef signrxs, signrys, signrzs, signrxyz, sign;
- LLVMValueRef arxs, arys, arzs;
- LLVMValueRef arx_ge_ary, maxarxsarys, arz_ge_arx_ary;
+ LLVMValueRef signs, signt, signr, signma;
+ LLVMValueRef as, at, ar;
+ LLVMValueRef as_ge_at, maxasat, ar_ge_as_at;
LLVMValueRef snewx, tnewx, snewy, tnewy, snewz, tnewz;
- LLVMValueRef ryneg, rzneg;
- LLVMValueRef ma, ima;
+ LLVMValueRef tnegi, rnegi;
+ LLVMValueRef ma, mai, ima;
LLVMValueRef posHalf = lp_build_const_vec(gallivm, coord_bld->type, 0.5);
LLVMValueRef signmask = lp_build_const_int_vec(gallivm, intctype,
1 << (intctype.width - 1));
LLVMValueRef facex = lp_build_const_int_vec(gallivm, intctype, PIPE_TEX_FACE_POS_X);
LLVMValueRef facey = lp_build_const_int_vec(gallivm, intctype, PIPE_TEX_FACE_POS_Y);
LLVMValueRef facez = lp_build_const_int_vec(gallivm, intctype, PIPE_TEX_FACE_POS_Z);
+ LLVMValueRef dmax[3], dmaxsnew, dmaxtnew;
assert(PIPE_TEX_FACE_NEG_X == PIPE_TEX_FACE_POS_X + 1);
assert(PIPE_TEX_FACE_NEG_Y == PIPE_TEX_FACE_POS_Y + 1);
assert(PIPE_TEX_FACE_NEG_Z == PIPE_TEX_FACE_POS_Z + 1);
- rx = LLVMBuildBitCast(builder, s, lp_build_vec_type(gallivm, intctype), "");
- ry = LLVMBuildBitCast(builder, t, lp_build_vec_type(gallivm, intctype), "");
- rz = LLVMBuildBitCast(builder, r, lp_build_vec_type(gallivm, intctype), "");
- ryneg = LLVMBuildXor(builder, ry, signmask, "");
- rzneg = LLVMBuildXor(builder, rz, signmask, "");
+ /*
+ * TODO do this only when needed, and implement explicit derivs (trivial).
+ */
+ if (need_derivs && implicit_derivs) {
+ LLVMValueRef ddx_ddy[2], tmp[2];
+ /*
+ * This isn't quite the same as the "ordinary" path since
+ * we need to extract the ds/dt/dr values before further processing.
+ */
+ static const unsigned char swizzle11[] = { /* no-op swizzle */
+ 0, LP_BLD_SWIZZLE_DONTCARE,
+ LP_BLD_SWIZZLE_DONTCARE, LP_BLD_SWIZZLE_DONTCARE
+ };
+ static const unsigned char swizzle12[] = {
+ 2, LP_BLD_SWIZZLE_DONTCARE,
+ LP_BLD_SWIZZLE_DONTCARE, LP_BLD_SWIZZLE_DONTCARE
+ };
+ static const unsigned char swizzle21[] = { /* no-op swizzle */
+ 0, LP_BLD_SWIZZLE_DONTCARE,
+ 2, LP_BLD_SWIZZLE_DONTCARE
+ };
+ static const unsigned char swizzle22[] = {
+ 1, LP_BLD_SWIZZLE_DONTCARE,
+ 3, LP_BLD_SWIZZLE_DONTCARE
+ };
+
+ ddx_ddy[0] = lp_build_packed_ddx_ddy_twocoord(coord_bld, s, t);
+ ddx_ddy[1] = lp_build_packed_ddx_ddy_onecoord(coord_bld, r);
+ ddx_ddy[0] = lp_build_abs(coord_bld, ddx_ddy[0]);
+ ddx_ddy[1] = lp_build_abs(coord_bld, ddx_ddy[1]);
+
+ tmp[0] = lp_build_swizzle_aos(coord_bld, ddx_ddy[0], swizzle21);
+ tmp[1] = lp_build_swizzle_aos(coord_bld, ddx_ddy[0], swizzle22);
+ dmax[0] = lp_build_max(coord_bld, tmp[0], tmp[1]);
+ dmax[1] = lp_build_swizzle_aos(coord_bld, dmax[0], swizzle12);
+
+ tmp[0] = lp_build_swizzle_aos(coord_bld, ddx_ddy[1], swizzle11);
+ tmp[1] = lp_build_swizzle_aos(coord_bld, ddx_ddy[1], swizzle12);
+ dmax[2] = lp_build_max(coord_bld, tmp[0], tmp[1]);
+ }
+ else if (need_derivs) {
+ /* dmax[0] = lp_build_max(coord_bld, derivs->ddx[0], derivs->ddy[0]);
+ dmax[1] = lp_build_max(coord_bld, derivs->ddx[1], derivs->ddy[1]);
+ dmax[2] = lp_build_max(coord_bld, derivs->ddx[2], derivs->ddy[2]); */
+ }
- /* the sign bit comes from the averaged vector (per quad),
- * as does the decision which face to use */
- signrxyz = LLVMBuildBitCast(builder, rxyz, lp_build_vec_type(gallivm, intctype), "");
- signrxyz = LLVMBuildAnd(builder, signrxyz, signmask, "");
+ si = LLVMBuildBitCast(builder, s, lp_build_vec_type(gallivm, intctype), "");
+ ti = LLVMBuildBitCast(builder, t, lp_build_vec_type(gallivm, intctype), "");
+ ri = LLVMBuildBitCast(builder, r, lp_build_vec_type(gallivm, intctype), "");
- arxs = lp_build_swizzle_scalar_aos(coord_bld, arxyz, 0);
- arys = lp_build_swizzle_scalar_aos(coord_bld, arxyz, 1);
- arzs = lp_build_swizzle_scalar_aos(coord_bld, arxyz, 2);
+ /*
+ * get absolute value (for x/y/z face selection) and sign bit
+ * (for mirroring minor coords and pos/neg face selection)
+ * of the original coords.
+ */
+ as = lp_build_abs(&bld->coord_bld, s);
+ at = lp_build_abs(&bld->coord_bld, t);
+ ar = lp_build_abs(&bld->coord_bld, r);
+ signs = LLVMBuildAnd(builder, si, signmask, "");
+ signt = LLVMBuildAnd(builder, ti, signmask, "");
+ signr = LLVMBuildAnd(builder, ri, signmask, "");
/*
- * select x if x >= y else select y
+ * major face determination: select x if x >= y else select y
* select previous result if y >= max(x,y) else select z
*/
- arx_ge_ary = lp_build_cmp(coord_bld, PIPE_FUNC_GEQUAL, arxs, arys);
- maxarxsarys = lp_build_max(coord_bld, arxs, arys);
- arz_ge_arx_ary = lp_build_cmp(coord_bld, PIPE_FUNC_GEQUAL, maxarxsarys, arzs);
+ as_ge_at = lp_build_cmp(coord_bld, PIPE_FUNC_GEQUAL, as, at);
+ maxasat = lp_build_max(coord_bld, as, at);
+ ar_ge_as_at = lp_build_cmp(coord_bld, PIPE_FUNC_GEQUAL, maxasat, ar);
/*
* compute all possible new s/t coords
- * snewx = signrx * -rz;
- * tnewx = -ry;
- * snewy = rx;
- * tnewy = signry * rz;
- * snewz = signrz * rx;
- * tnewz = -ry;
+ * snewx = signs * -r;
+ * tnewx = -t;
+ * snewy = s;
+ * tnewy = signt * r;
+ * snewz = signr * s;
+ * tnewz = -t;
*/
- signrxs = lp_build_swizzle_scalar_aos(cint_bld, signrxyz, 0);
- snewx = LLVMBuildXor(builder, signrxs, rzneg, "");
- tnewx = ryneg;
+ tnegi = LLVMBuildXor(builder, ti, signmask, "");
+ rnegi = LLVMBuildXor(builder, ri, signmask, "");
+
+ snewx = LLVMBuildXor(builder, signs, rnegi, "");
+ tnewx = tnegi;
- signrys = lp_build_swizzle_scalar_aos(cint_bld, signrxyz, 1);
- snewy = rx;
- tnewy = LLVMBuildXor(builder, signrys, rz, "");
+ snewy = si;
+ tnewy = LLVMBuildXor(builder, signt, ri, "");
- signrzs = lp_build_swizzle_scalar_aos(cint_bld, signrxyz, 2);
- snewz = LLVMBuildXor(builder, signrzs, rx, "");
- tnewz = ryneg;
+ snewz = LLVMBuildXor(builder, signr, si, "");
+ tnewz = tnegi;
/* XXX on x86 unclear if we should cast the values back to float
* or not - on some cpus (nehalem) pblendvb has twice the throughput
* transition penalties when using it (this depends on what llvm
* will chose for the bit ops above so there appears no "right way",
* but given the boatload of selects let's just use the int type).
- *
- * Unfortunately we also need the sign bit of the summed coords.
*/
- *face_s = lp_build_select(cint_bld, arx_ge_ary, snewx, snewy);
- *face_t = lp_build_select(cint_bld, arx_ge_ary, tnewx, tnewy);
- ma = lp_build_select(coord_bld, arx_ge_ary, s, t);
- *face = lp_build_select(cint_bld, arx_ge_ary, facex, facey);
- sign = lp_build_select(cint_bld, arx_ge_ary, signrxs, signrys);
-
- *face_s = lp_build_select(cint_bld, arz_ge_arx_ary, *face_s, snewz);
- *face_t = lp_build_select(cint_bld, arz_ge_arx_ary, *face_t, tnewz);
- ma = lp_build_select(coord_bld, arz_ge_arx_ary, ma, r);
- *face = lp_build_select(cint_bld, arz_ge_arx_ary, *face, facez);
- sign = lp_build_select(cint_bld, arz_ge_arx_ary, sign, signrzs);
+
+ /* select/mirror */
+ *face_s = lp_build_select(cint_bld, as_ge_at, snewx, snewy);
+ *face_t = lp_build_select(cint_bld, as_ge_at, tnewx, tnewy);
+ ma = lp_build_select(coord_bld, as_ge_at, s, t);
+ *face = lp_build_select(cint_bld, as_ge_at, facex, facey);
+ if (need_derivs) {
+ dmaxsnew = lp_build_select(coord_bld, as_ge_at, dmax[2], dmax[0]);
+ dmaxtnew = lp_build_select(coord_bld, as_ge_at, dmax[1], dmax[2]);
+ }
+
+ *face_s = lp_build_select(cint_bld, ar_ge_as_at, *face_s, snewz);
+ *face_t = lp_build_select(cint_bld, ar_ge_as_at, *face_t, tnewz);
+ ma = lp_build_select(coord_bld, ar_ge_as_at, ma, r);
+ *face = lp_build_select(cint_bld, ar_ge_as_at, *face, facez);
+ if (need_derivs) {
+ dmaxsnew = lp_build_select(coord_bld, ar_ge_as_at, dmaxsnew, dmax[0]);
+ dmaxtnew = lp_build_select(coord_bld, ar_ge_as_at, dmaxtnew, dmax[1]);
+ }
*face_s = LLVMBuildBitCast(builder, *face_s,
lp_build_vec_type(gallivm, coord_bld->type), "");
* as long as we ensure vblendvps gets used we can actually
* skip the comparison and just use sign as a "mask" directly.
*/
- sign = LLVMBuildLShr(builder, sign, signshift, "");
- *face = LLVMBuildOr(builder, *face, sign, "face");
+ mai = LLVMBuildBitCast(builder, ma, lp_build_vec_type(gallivm, intctype), "");
+ signma = LLVMBuildLShr(builder, mai, signshift, "");
+ *face = LLVMBuildOr(builder, *face, signma, "face");
ima = lp_build_cube_imapos(coord_bld, ma);
+ /* project coords */
*face_s = lp_build_mul(coord_bld, *face_s, ima);
*face_s = lp_build_add(coord_bld, *face_s, posHalf);
*face_t = lp_build_mul(coord_bld, *face_t, ima);
*face_t = lp_build_add(coord_bld, *face_t, posHalf);
+
+ /* project derivs */
+ if (need_derivs) {
+ /*
+ * we do some optimization here, since we know it's square
+ * we can do the max before projection (and before size mul,
+ * which the so-called "rho" is missing here).
+ * For explicit derivs this is fully per-pixel vector, for implicit
+ * derivs only the first value per quad contains useful values.
+ */
+ *rho = lp_build_max(coord_bld, dmaxsnew, dmaxtnew);
+ *rho = lp_build_mul(coord_bld, *rho, ima);
+ }
}
else {
LLVMValueRef shuffles[4];
LLVMValueRef arxy_ge_aryx, arxy_ge_arzz, arxy_ge_arxy_arzz;
LLVMValueRef arxyxy, aryxzz, arxyxy_ge_aryxzz;
+ LLVMValueRef tmp[4], rxyz, arxyz;
struct lp_build_context *float_bld = &bld->float_bld;
assert(bld->coord_bld.type.length == 4);
+ tmp[0] = s;
+ tmp[1] = t;
+ tmp[2] = r;
+ rxyz = lp_build_hadd_partial4(&bld->coord_bld, tmp, 3);
+ arxyz = lp_build_abs(&bld->coord_bld, rxyz);
+
shuffles[0] = lp_build_const_int32(gallivm, 0);
shuffles[1] = lp_build_const_int32(gallivm, 1);
shuffles[2] = lp_build_const_int32(gallivm, 0);
{
/* +/- X face */
LLVMValueRef sign, ima;
- rx = LLVMBuildExtractElement(builder, rxyz,
+ si = LLVMBuildExtractElement(builder, rxyz,
lp_build_const_int32(gallivm, 0), "");
/* +/- X face */
- sign = lp_build_sgn(float_bld, rx);
+ sign = lp_build_sgn(float_bld, si);
ima = lp_build_cube_imaneg(coord_bld, s);
*face_s = lp_build_cube_coord(coord_bld, sign, +1, r, ima);
*face_t = lp_build_cube_coord(coord_bld, NULL, +1, t, ima);
- *face = lp_build_cube_face(bld, rx,
+ *face = lp_build_cube_face(bld, si,
PIPE_TEX_FACE_POS_X,
PIPE_TEX_FACE_NEG_X);
LLVMBuildStore(builder, *face_s, face_s_var);
{
LLVMValueRef sign, ima;
/* +/- Y face */
- ry = LLVMBuildExtractElement(builder, rxyz,
+ ti = LLVMBuildExtractElement(builder, rxyz,
lp_build_const_int32(gallivm, 1), "");
- sign = lp_build_sgn(float_bld, ry);
+ sign = lp_build_sgn(float_bld, ti);
ima = lp_build_cube_imaneg(coord_bld, t);
*face_s = lp_build_cube_coord(coord_bld, NULL, -1, s, ima);
*face_t = lp_build_cube_coord(coord_bld, sign, -1, r, ima);
- *face = lp_build_cube_face(bld, ry,
+ *face = lp_build_cube_face(bld, ti,
PIPE_TEX_FACE_POS_Y,
PIPE_TEX_FACE_NEG_Y);
LLVMBuildStore(builder, *face_s, face_s_var);
{
/* +/- Z face */
LLVMValueRef sign, ima;
- rz = LLVMBuildExtractElement(builder, rxyz,
+ ri = LLVMBuildExtractElement(builder, rxyz,
lp_build_const_int32(gallivm, 2), "");
- sign = lp_build_sgn(float_bld, rz);
+ sign = lp_build_sgn(float_bld, ri);
ima = lp_build_cube_imaneg(coord_bld, r);
*face_s = lp_build_cube_coord(coord_bld, sign, -1, s, ima);
*face_t = lp_build_cube_coord(coord_bld, NULL, +1, t, ima);
- *face = lp_build_cube_face(bld, rz,
+ *face = lp_build_cube_face(bld, ri,
PIPE_TEX_FACE_POS_Z,
PIPE_TEX_FACE_NEG_Z);
LLVMBuildStore(builder, *face_s, face_s_var);